Knowledge Extraction from Trained ANN River Water Quality Parameters

Environmental factors, agricultural production and human health are impacted by the evaluation of water quality indicators: chemical oxygen demand (COD), Dissolved Oxygen (DO) and biological oxygen demand (BOD). Artificial Neural Networks (ANN) have recently made it possible to automate the assessment of water quality parameters with a faster and more accurate evaluation. The current study aims to develop a water quality model for the critical parameters: DO-BOD-COD for the Mula-Mutha River, Pune India, by using ANN. ANNs can simulate non-linear relationships between input and output parameters in water quality assessment. Even though ANN has proven to be highly effective in several cases, many scholars are still hesitant to accept ANN in practise as it is labelled as "black box". This skeptical thinking arises from the lack of in-depth description regarding the decision-making process that ANN uses to determine the relationship between input and output. The primary goal of this work is to employ a Knowledge Extraction (KE) approach to retrieve knowledge that is locked in the trained biases and weights of the trained DO-BOD-COD models by ANN. The approach developed by Shah and Londhe in 2017 is employed for the process. By accurately evaluating the effect of the input parameters on the water quality parameter as directly or inversely proportional along with the magnitude of influence; is supported by the physics of the underlying process. The study shows EC, Alk and DO as influential parameters for COD, DO and BOD prediction which is as per the comprehension of physics.